Rapid Screening of Chemicals with Placental Transfer Risk Using Interpretable Machine Learning

Assessing the placental transfer efficiency of toxic chemicals remains challenging. Here, a robust machine learning (ML) model was developed to predict the human fetal–maternal blood concentration ratio (F/M) at the exposomic level. By curating one of the largest F/M data sets, we evaluated a series of prediction models using a combination of 12 ML algorithms and four molecular fingerprints. The long short-term memory (LSTM) model with retraining optimization works as the best performer, displayed robust accuracy (R²_train = 0.91, R²_test = 0.68), and was subsequently applied to our previously developed risk-based Human Exposome and Metabolite Database (HExpMetDB). The fetal hazard quotient (FHQ) was assessed using the predicted F/M ratios, probabilistic exposure dose, and toxicity index. From the top 1000 prioritized chemicals via FHQs ranking, we randomly selected four candidates (triethyl phosphate, benzotriazole, oxybenzone, and dichlormid) to perform in vivo experiments. All four chemicals exhibited transplacental potential (F/M ratio >0.3) as new possible chemicals of concern, demonstrating the accuracy of the predictive model. The Shapley additive explanation (SHAP) method revealed the top 10 key structural fragments related to the transplacental transfer efficiency. We believe that the prediction model can serve as an effective tool to screen potential risk compounds of fetal exposure.